qnite labs

Qnite LabsDREAM × CACHE · PGK2 · 2026

First attempt.
Two weeks.
Top 10 of 110.

2weeks
to build
7.5M
molecules scored
17 / 39
known binders in our top 1%

Mohammad Afshari · Erfan Ebrahimi · Erfan Aalivand · Arshia Bordbar

The resultFinal leaderboard

Two submissions.
5th and 12th of 110.

51015Chemical series found · top 20 of 1101#1 · AIIF15 series · 19 hits2345#5 · Er-test_v112 series · 28 hits6789101112#12 · StableFold_AL210 series · 27 hits1314151617181920rank

Ranked by chemical series found, then hits. Source: Synapse leaderboard, 3 Nov 2026.

Target 2035

The arenaTarget 2035

A chemical probe
for every human protein.
By 2035.

~20,000
human proteins in scope
200+
chemical probes developed by SGC
4,000+
protein structures made public

Sources: SGC Backgrounder (Nov 2025); ASBMB Today (2021)

BayerBoehringer IngelheimPfizerTakedaGates Foundation
National Institutes of HealthMerck KGaAJANSSENBristol Myers SquibbGenentech
FAPESPMICHAEL J. FOXFOUNDATIONEU INNOVATIVEHEALTH INITIATIVECANADA FOUNDATIONFOR INNOVATION

The arenaWho backs it

Pharma, foundations,
governments.
One open dataset.

A global open-science effort led by the Structural Genomics Consortium to find a chemical probe for every human protein by 2035. Its open DEL and mass-spec data now feed machine-learning challenges like ours.

They stand behind the competition, not behind Qnite Labs.

Sources: SGC Backgrounders (2023, Nov 2025)

The arenaThe exam and the referee

Teams predict.
A lab decides.
Everything is public.

CACHE · the exam

8
challenges since Dec 2021
23–24
teams per challenge
1.5–1.7K
compounds lab-tested each

DREAM · the referee

60+
public challenges
30,000+
solvers
105+
publications

Sources: CACHE results pages (#3–#5); Conscience (2023); Sage Bionetworks; ISCB RSG/DREAM

The arenaOur challenge

One target: PGK2.
A billion molecules.
Two phases.

  1. PHASE 1First roundRank 24 with a team of two.
  2. PHASE 2Active learningRanked by chemical series, then hits. Top 10 of 110 with a team of four.

Sources: CACHE challenge page; Synapse leaderboard, 3 Nov 2026

The challengeDNA-encoded libraries

Three decades of DEL.
Still an open problem.

1992The ideaBrenner & Lerner2004Bead-free DELsLiu · Neri · Harbury2009800M in one libraryGSK2015Screening live cellsGSK · 15B compounds2017Noise bitesbig libraries lose binders2020Machine learningGoogle + X-Chem · 30% hits2024Pharma pools its DELsPfizer + peers2026Open dataTarget 2035 · AIRCHECK

GSK today: 100+ libraries, 3.5 billion structures, 10 clinical candidates, and still fighting noisy data and overfitting.

Sources: Nature Rev Drug Discov (2016); PMC8369695; J Med Chem (McCloskey, 2020); GSK talk, J. Messer (2025); Pfizer (2024); SGC and AIRCHECK (2026)

The challengeWhy it is so hard

Real signal is a sliver
inside a lot of noise.

Share of each library that is real signalqDOS140.08%qDOS19_10.01%qDOS240.02%qDOS19_20.05%qDOS9<0.01%qDOS230.84%qDOS110.04%qDOS42_20.25%qDOS300.4%qDOS18_30.09%Green: passed our evidence gate1Yields varyno reaction runs to completion2Truncates look identicalsame DNA tag as the real product3PCR and sequencing add noiselow counts are mostly luck4Counts are not affinityonly resynthesis confirms a hit

Sources: Komar & Kalinić (2020); Lim et al. (2022); Wichert et al., Chem Rev (2024); Menzeleev et al. (2026)

Arm 1 · DEL modelThe problem

Almost all of it
is noise.

98.96%read exactly once1.04%read more than once · where every real hit livesLibrary qDOS30: 1,709,679 of 1,727,652 molecules

Arm 1 · DEL modelThe idea

Don’t guess a cutoff.
Derive it.

0%50%100%real if 50% or more0.2%1 read14%2 reads93%3 readsthe cutoff99.9%4 readsChance a molecule is real signal, by read count. Crossing at 3 reads in nine libraries, 4 in one.

Arm 1 · DEL modelThe proof

Real binders,
chemistry it never saw.

0%50%100%top 0.1%top 1%top 10%allchance3/3917/3929/39184,632 compounds from a different assay. 17 of 39 known binders in the top 1%, 44× chance.share of the ranked list examined

Arm 1 · DEL modelWhere it breaks

We publish the failure,
not just the win.

0.500.751.000.5 = coin flip0.962Random split0.959Unseen BB10.827Unseen BB20.827Unseen BB3Pooled AUC when whole building blocks are held out. Worst single library: 0.586.

Arm 1 · DEL modelHow it is built · click to advance

The DEL model: six build stages, three checks.

Read counts — A DEL screen gives no labels, only how often each compound was read.Read counts10 librariesMixture fit — Per library, fit a mixture: a huge noise background plus a small signal tail.1Mixture fitnoise vs signalLabels — Each read count gets a probability of being real. 0.5 or more is a positive.2Labelsfrom a posteriorFair folds — Folds are cut by building-block pair, so close relatives never straddle train and test.3Fair foldsno leakageBake-off — Three rounds under fair folds: keep all negatives, ECFP4 fingerprints, LightGBM.4Bake-off3 roundsGate — Keep a library only with 100+ positives and a held-out AUC bound of 0.60+.5Gate5 of 10 passCross-fit — Each fold-model scores only compounds it never saw. Calibration and library weights give one ranking.6Cross-fitcalibrate + rankEnrichment — In the top 1% of each library: 28 to 81 times more signal than chance.Enrichmenttop 1% of libraryHard split — Hold out every compound with a given building block and accuracy falls.Hard splitunseen building blockBlind test — On 184,632 compounds from mass spectrometry, 17 of 39 known binders land in our top 1%.Blind testa different assay
Input
7.5M

Read counts

A DEL screen gives no labels, only how often each compound was read.

KEY NUMBERS

Most were read once. Real binders hide in a thin tail.

The strategy in six stages
The strategy in six stages
Build stage 1 of 6
noise + signal

Fit the assay

Per library, fit a mixture: a huge noise background plus a small signal tail.

KEY NUMBERS

Counts only, no chemistry yet.

A DEL library is a huge background plus a thin signal tail
A DEL library is a huge background plus a thin signal tail
Build stage 2 of 6
3 reads

Labels from a posterior

Each read count gets a probability of being real. 0.5 or more is a positive.

KEY NUMBERS

Cutoff: 3 reads in nine libraries, 4 in one. 16,238 positives.

Probability a compound is real signal, by read count
Probability a compound is real signal, by read count
What the labels look like: positives are a sliver
What the labels look like: positives are a sliver
Build stage 3 of 6
no leakage

Fair folds

Folds are cut by building-block pair, so close relatives never straddle train and test.

WHY IT MATTERS

Every number we report respects this.

Random split vs building-block split
Random split vs building-block split
Build stage 4 of 6
AUC 0.96

Bake-off

Three rounds under fair folds: keep all negatives, ECFP4 fingerprints, LightGBM.

KEY NUMBERS

Cleaning the negatives made it worse.

The three-round bake-off
The three-round bake-off
The model is 400 trees asking: is this bit ON?
The model is 400 trees asking: is this bit ON?
Build stage 5 of 6
5 of 10

Evidence gate

Keep a library only with 100+ positives and a held-out AUC bound of 0.60+.

KEY NUMBERS

qDOS14: AUC 0.91 on 67 positives, dropped.

The gate drops 5 of 10 libraries
The gate drops 5 of 10 libraries
Build stage 6 of 6
7.5M scored

Cross-fit and rank

Each fold-model scores only compounds it never saw. Calibration and library weights give one ranking.

KEY NUMBERS

A ranking score, not a probability.

Cross-fitted scores: positives right, background left
Cross-fitted scores: positives right, background left
One molecule, end to end
One molecule, end to end
Check 1 of 3
28–81×

Enrichment

In the top 1% of each library: 28 to 81 times more signal than chance.

KEY NUMBERS

Top 0.1%: 54 to 324 times.

Enrichment on the full population
Enrichment on the full population
Check 2 of 3
0.96 → 0.83

Hard split

Hold out every compound with a given building block and accuracy falls.

WHAT IT TELLS US

One library drops to 0.59, near a coin flip.

Hold out whole building blocks
Hold out whole building blocks
Check 3 of 3
17 / 39

A different assay

On 184,632 compounds from mass spectrometry, 17 of 39 known binders land in our top 1%.

KEY NUMBERS

44× chance. The labelled block sits higher than average.

Where the 39 known binders land
Where the 39 known binders land

Click any node to see its chart in the side panel. Click a chart to enlarge it.

Arm 2 · SelectorFilter, find, cluster · click to advance

The selector: builds on the DEL filter.

Test pool — The compounds we may submit. We must choose 50.Test pool178,052 submittable39 known hits — A random forest on ECFP4 fingerprints learns from the 6,580 labelled compounds and re-ranks what the filter kept.39 known hits6,580 labelledDEL model — Trained on DEL read counts, with no hit labels. It scores any molecule and shrinks the pool.1DEL modelthe filterSelector — A random forest on ECFP4 fingerprints learns from the 6,580 labelled compounds and re-ranks what the filter kept.2Selectorfinds the bestComplement — The rankings disagree across the list, agree at the top, and explore different chemistry.Complementwhy bothCluster the tops — Group the top of the list into clusters of similar compounds and take one from each.3Cluster the topsone per cluster50 picks — The selector with this clustering rule scored 28 hits in 12 series: 5th of 110.50 picksmany series
Input
178,052

The test pool

The compounds we may submit. We must choose 50.

KEY NUMBERS

Ranking counts chemical series first, then hits.

WDR91 leaderboard: novelty beat raw hit count
WDR91 leaderboard: novelty beat raw hit count
Stage 1 of 3
17 / 39

DEL model: the filter

Trained on DEL read counts, with no hit labels. It scores any molecule and shrinks the pool.

KEY NUMBERS

Its top 1% already holds 17 of the 39 known binders. Its top 10% holds 29.

Where the 39 known binders land
Where the 39 known binders land
Stage 2 of 3
39 hits

Selector: find the best

A random forest on ECFP4 fingerprints learns from the 6,580 labelled compounds and re-ranks what the filter kept.

KEY NUMBERS

With a whole chemotype held out, 46% of hidden hits reach its top 1%, none its top 50.

Why both
ρ 0.13

They complete each other

The rankings disagree across the list, agree at the top, and explore different chemistry.

KEY NUMBERS

108 shared in the top 1% against 18 by chance. Scaffolds in the top 1000: 777 (DEL) vs 548 (selector).

DEL model vs selector: two different rankings
DEL model vs selector: two different rankings
Scaffolds in the top 1000: DEL model vs selector
Scaffolds in the top 1000: DEL model vs selector
Stage 3 of 3
1 / cluster

Cluster the tops

Group the top of the list into clusters of similar compounds and take one from each.

KEY NUMBERS

Rule used: top 3,000, Butina clusters at 0.6 similarity, one pick per cluster.

Result
28 hits

50 picks, 12 series

The selector with this clustering rule scored 28 hits in 12 series: 5th of 110.

WHAT IT TELLS US

46 of 50 picks sit in series that already hold a known hit.

Where the 50 picks sit next to the 39 known hits
Where the 50 picks sit next to the 39 known hits

Click any node to see its note and chart in the side panel.

Qnite LabsWhat we proved

Same system.
More expertise. Top 10.

PHASE 1#24Team of twoPHASE 2Top 10Team of fourPharmacistChemistthe same multi-agent system+ domain expertise · 2 weeks

Phase 1: a multi-agent AI system reached rank 24 with two people. Phase 2: a pharmacist and a chemist joined, the system stayed the same.

← → or click to navigate · F for fullscreen